fix(metal): revert inactive NAX MoE simdgroup skipping

[jonathan308]

Summary

Revert #4352's inactive-simdgroup NAX MoE optimization after a physical mixed-Mac tensor-parallel A/B isolated it as the trigger for repeatable lost JACCL completions.

Evidence

• v0.32.2 with Skip unnecessary simdgroup computations for quantised MOE matmuls on NAX #4352: two clean 30K DS4 TP2 failures, including one after reboot; rank 1 stops making all_reduce progress and the request returns zero tokens.
• v0.32.2 with only Skip unnecessary simdgroup computations for quantised MOE matmuls on NAX #4352 reverted: the original 120K -> 30K failure sequence passes, plus an additional cold 30K rotation, cache reuse, B2 decode, and mixed decode+prefill.
• 3,440 dependent synthetic collectives complete correctly in 0.829 seconds on the reverted build.

The detailed physical reproducer and measurements are linked in the issue.

Why a revert

The issue is hardware/workload-specific and current CI cannot exercise a heterogeneous M3 Ultra + M5 Max JACCL pair. Restoring the pre-#4352 code is the smallest proven-safe change. A follow-up can reintroduce the optimization with a safe uniform execution scheme or an explicit capability gate.

This PR is a draft pending maintainer guidance and any additional isolated-kernel reproducer they want.

Refs #4403.

[nastya236]

Thanks for flagging this. I will look into it soon.

[zcbenz]

/cc @RohanGautam